@Test
// All Task attempts are timed out, leading to Job failure
public void testTimedOutTask() throws Exception {
MRApp app = new TimeOutTaskMRApp(1, 0);
Configuration conf = new Configuration();
int maxAttempts = 2;
conf.setInt(MRJobConfig.MAP_MAX_ATTEMPTS, maxAttempts);
// disable uberization (requires entire job to be reattempted, so max for
// subtask attempts is overridden to 1)
conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
Job job = app.submit(conf);
app.waitForState(job, JobState.FAILED);
Map<TaskId, Task> tasks = job.getTasks();
Assert.assertEquals("Num tasks is not correct", 1, tasks.size());
Task task = tasks.values().iterator().next();
Assert.assertEquals(
"Task state not correct", TaskState.FAILED, task.getReport().getTaskState());
Map<TaskAttemptId, TaskAttempt> attempts = tasks.values().iterator().next().getAttempts();
Assert.assertEquals("Num attempts is not correct", maxAttempts, attempts.size());
for (TaskAttempt attempt : attempts.values()) {
Assert.assertEquals(
"Attempt state not correct",
TaskAttemptState.FAILED,
attempt.getReport().getTaskAttemptState());
}
}
@Test
public void testCommitPending() throws Exception {
MRApp app = new MRApp(1, 0, false, this.getClass().getName(), true);
Job job = app.submit(new Configuration());
app.waitForState(job, JobState.RUNNING);
Assert.assertEquals("Num tasks not correct", 1, job.getTasks().size());
Iterator<Task> it = job.getTasks().values().iterator();
Task task = it.next();
app.waitForState(task, TaskState.RUNNING);
TaskAttempt attempt = task.getAttempts().values().iterator().next();
app.waitForState(attempt, TaskAttemptState.RUNNING);
// send the commit pending signal to the task
app.getContext()
.getEventHandler()
.handle(new TaskAttemptEvent(attempt.getID(), TaskAttemptEventType.TA_COMMIT_PENDING));
// wait for first attempt to commit pending
app.waitForState(attempt, TaskAttemptState.COMMIT_PENDING);
// send the done signal to the task
app.getContext()
.getEventHandler()
.handle(
new TaskAttemptEvent(
task.getAttempts().values().iterator().next().getID(),
TaskAttemptEventType.TA_DONE));
app.waitForState(job, JobState.SUCCEEDED);
}
@Override
public void updateAttempt(TaskAttemptStatus status, long timestamp) {
TaskAttemptId attemptID = status.id;
TaskId taskID = attemptID.getTaskId();
JobId jobID = taskID.getJobId();
Job job = context.getJob(jobID);
if (job == null) {
return;
}
Task task = job.getTask(taskID);
if (task == null) {
return;
}
Long boxedStart = startTimes.get(attemptID);
long start = boxedStart == null ? Long.MIN_VALUE : boxedStart;
TaskAttempt taskAttempt = task.getAttempt(attemptID);
if (taskAttempt.getState() == TaskAttemptState.SUCCEEDED) {
boolean isNew = false;
// is this a new success?
synchronized (doneTasks) {
if (!doneTasks.contains(task)) {
doneTasks.add(task);
isNew = true;
}
}
// It's a new completion
// Note that if a task completes twice [because of a previous speculation
// and a race, or a success followed by loss of the machine with the
// local data] we only count the first one.
if (isNew) {
long finish = timestamp;
if (start > 1L && finish > 1L && start <= finish) {
long duration = finish - start;
DataStatistics statistics = dataStatisticsForTask(taskID);
if (statistics != null) {
statistics.add(duration);
}
}
}
}
}
@Test(timeout = 10000)
public void testFailAbortDoesntHang() throws IOException {
Configuration conf = new Configuration();
conf.set(MRJobConfig.MR_AM_STAGING_DIR, stagingDir);
conf.set(MRJobConfig.MR_AM_COMMITTER_CANCEL_TIMEOUT_MS, "1000");
DrainDispatcher dispatcher = new DrainDispatcher();
dispatcher.init(conf);
dispatcher.start();
OutputCommitter committer = Mockito.mock(OutputCommitter.class);
CommitterEventHandler commitHandler = createCommitterEventHandler(dispatcher, committer);
commitHandler.init(conf);
commitHandler.start();
// Job has only 1 mapper task. No reducers
conf.setInt(MRJobConfig.NUM_REDUCES, 0);
conf.setInt(MRJobConfig.MAP_MAX_ATTEMPTS, 1);
JobImpl job = createRunningStubbedJob(conf, dispatcher, 1, null);
// Fail / finish all the tasks. This should land the JobImpl directly in the
// FAIL_ABORT state
for (Task t : job.tasks.values()) {
TaskImpl task = (TaskImpl) t;
task.handle(new TaskEvent(task.getID(), TaskEventType.T_SCHEDULE));
for (TaskAttempt ta : task.getAttempts().values()) {
task.handle(new TaskTAttemptEvent(ta.getID(), TaskEventType.T_ATTEMPT_FAILED));
}
}
assertJobState(job, JobStateInternal.FAIL_ABORT);
dispatcher.await();
// Verify abortJob is called once and the job failed
Mockito.verify(committer, Mockito.timeout(2000).times(1))
.abortJob((JobContext) Mockito.any(), (State) Mockito.any());
assertJobState(job, JobStateInternal.FAILED);
dispatcher.stop();
}
@Test
public void testTaskFailWithUnusedContainer() throws Exception {
MRApp app = new MRAppWithFailingTaskAndUnusedContainer();
Configuration conf = new Configuration();
int maxAttempts = 1;
conf.setInt(MRJobConfig.MAP_MAX_ATTEMPTS, maxAttempts);
// disable uberization (requires entire job to be reattempted, so max for
// subtask attempts is overridden to 1)
conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
Job job = app.submit(conf);
app.waitForState(job, JobState.RUNNING);
Map<TaskId, Task> tasks = job.getTasks();
Assert.assertEquals("Num tasks is not correct", 1, tasks.size());
Task task = tasks.values().iterator().next();
app.waitForState(task, TaskState.SCHEDULED);
Map<TaskAttemptId, TaskAttempt> attempts = tasks.values().iterator().next().getAttempts();
Assert.assertEquals("Num attempts is not correct", maxAttempts, attempts.size());
TaskAttempt attempt = attempts.values().iterator().next();
app.waitForInternalState((TaskAttemptImpl) attempt, TaskAttemptStateInternal.ASSIGNED);
app.getDispatcher()
.getEventHandler()
.handle(new TaskAttemptEvent(attempt.getID(), TaskAttemptEventType.TA_CONTAINER_COMPLETED));
app.waitForState(job, JobState.FAILED);
}
@Test
public void testHistoryParsingForFailedAttempts() throws Exception {
LOG.info("STARTING testHistoryParsingForFailedAttempts");
try {
Configuration conf = new Configuration();
conf.setClass(
CommonConfigurationKeysPublic.NET_TOPOLOGY_NODE_SWITCH_MAPPING_IMPL_KEY,
MyResolver.class,
DNSToSwitchMapping.class);
RackResolver.init(conf);
MRApp app =
new MRAppWithHistoryWithFailedAttempt(2, 1, true, this.getClass().getName(), true);
app.submit(conf);
Job job = app.getContext().getAllJobs().values().iterator().next();
JobId jobId = job.getID();
app.waitForState(job, JobState.SUCCEEDED);
// make sure all events are flushed
app.waitForState(Service.STATE.STOPPED);
String jobhistoryDir = JobHistoryUtils.getHistoryIntermediateDoneDirForUser(conf);
JobHistory jobHistory = new JobHistory();
jobHistory.init(conf);
JobIndexInfo jobIndexInfo = jobHistory.getJobFileInfo(jobId).getJobIndexInfo();
String jobhistoryFileName = FileNameIndexUtils.getDoneFileName(jobIndexInfo);
Path historyFilePath = new Path(jobhistoryDir, jobhistoryFileName);
FSDataInputStream in = null;
FileContext fc = null;
try {
fc = FileContext.getFileContext(conf);
in = fc.open(fc.makeQualified(historyFilePath));
} catch (IOException ioe) {
LOG.info("Can not open history file: " + historyFilePath, ioe);
throw (new Exception("Can not open History File"));
}
JobHistoryParser parser = new JobHistoryParser(in);
JobInfo jobInfo = parser.parse();
Exception parseException = parser.getParseException();
Assert.assertNull("Caught an expected exception " + parseException, parseException);
int noOffailedAttempts = 0;
Map<TaskID, TaskInfo> allTasks = jobInfo.getAllTasks();
for (Task task : job.getTasks().values()) {
TaskInfo taskInfo = allTasks.get(TypeConverter.fromYarn(task.getID()));
for (TaskAttempt taskAttempt : task.getAttempts().values()) {
TaskAttemptInfo taskAttemptInfo =
taskInfo.getAllTaskAttempts().get(TypeConverter.fromYarn((taskAttempt.getID())));
// Verify rack-name for all task attempts
Assert.assertEquals("rack-name is incorrect", taskAttemptInfo.getRackname(), RACK_NAME);
if (taskAttemptInfo.getTaskStatus().equals("FAILED")) {
noOffailedAttempts++;
}
}
}
Assert.assertEquals("No of Failed tasks doesn't match.", 2, noOffailedAttempts);
} finally {
LOG.info("FINISHED testHistoryParsingForFailedAttempts");
}
}
private void checkHistoryParsing(
final int numMaps, final int numReduces, final int numSuccessfulMaps) throws Exception {
Configuration conf = new Configuration();
conf.set(MRJobConfig.USER_NAME, System.getProperty("user.name"));
long amStartTimeEst = System.currentTimeMillis();
conf.setClass(
CommonConfigurationKeysPublic.NET_TOPOLOGY_NODE_SWITCH_MAPPING_IMPL_KEY,
MyResolver.class,
DNSToSwitchMapping.class);
RackResolver.init(conf);
MRApp app = new MRAppWithHistory(numMaps, numReduces, true, this.getClass().getName(), true);
app.submit(conf);
Job job = app.getContext().getAllJobs().values().iterator().next();
JobId jobId = job.getID();
LOG.info("JOBID is " + TypeConverter.fromYarn(jobId).toString());
app.waitForState(job, JobState.SUCCEEDED);
// make sure all events are flushed
app.waitForState(Service.STATE.STOPPED);
String jobhistoryDir = JobHistoryUtils.getHistoryIntermediateDoneDirForUser(conf);
FileContext fc = null;
try {
fc = FileContext.getFileContext(conf);
} catch (IOException ioe) {
LOG.info("Can not get FileContext", ioe);
throw (new Exception("Can not get File Context"));
}
if (numMaps == numSuccessfulMaps) {
String summaryFileName = JobHistoryUtils.getIntermediateSummaryFileName(jobId);
Path summaryFile = new Path(jobhistoryDir, summaryFileName);
String jobSummaryString = getJobSummary(fc, summaryFile);
Assert.assertNotNull(jobSummaryString);
Assert.assertTrue(jobSummaryString.contains("resourcesPerMap=100"));
Assert.assertTrue(jobSummaryString.contains("resourcesPerReduce=100"));
Map<String, String> jobSummaryElements = new HashMap<String, String>();
StringTokenizer strToken = new StringTokenizer(jobSummaryString, ",");
while (strToken.hasMoreTokens()) {
String keypair = strToken.nextToken();
jobSummaryElements.put(keypair.split("=")[0], keypair.split("=")[1]);
}
Assert.assertEquals(
"JobId does not match", jobId.toString(), jobSummaryElements.get("jobId"));
Assert.assertEquals("JobName does not match", "test", jobSummaryElements.get("jobName"));
Assert.assertTrue(
"submitTime should not be 0", Long.parseLong(jobSummaryElements.get("submitTime")) != 0);
Assert.assertTrue(
"launchTime should not be 0", Long.parseLong(jobSummaryElements.get("launchTime")) != 0);
Assert.assertTrue(
"firstMapTaskLaunchTime should not be 0",
Long.parseLong(jobSummaryElements.get("firstMapTaskLaunchTime")) != 0);
Assert.assertTrue(
"firstReduceTaskLaunchTime should not be 0",
Long.parseLong(jobSummaryElements.get("firstReduceTaskLaunchTime")) != 0);
Assert.assertTrue(
"finishTime should not be 0", Long.parseLong(jobSummaryElements.get("finishTime")) != 0);
Assert.assertEquals(
"Mismatch in num map slots",
numSuccessfulMaps,
Integer.parseInt(jobSummaryElements.get("numMaps")));
Assert.assertEquals(
"Mismatch in num reduce slots",
numReduces,
Integer.parseInt(jobSummaryElements.get("numReduces")));
Assert.assertEquals(
"User does not match", System.getProperty("user.name"), jobSummaryElements.get("user"));
Assert.assertEquals("Queue does not match", "default", jobSummaryElements.get("queue"));
Assert.assertEquals("Status does not match", "SUCCEEDED", jobSummaryElements.get("status"));
}
JobHistory jobHistory = new JobHistory();
jobHistory.init(conf);
HistoryFileInfo fileInfo = jobHistory.getJobFileInfo(jobId);
JobInfo jobInfo;
long numFinishedMaps;
synchronized (fileInfo) {
Path historyFilePath = fileInfo.getHistoryFile();
FSDataInputStream in = null;
LOG.info("JobHistoryFile is: " + historyFilePath);
try {
in = fc.open(fc.makeQualified(historyFilePath));
} catch (IOException ioe) {
LOG.info("Can not open history file: " + historyFilePath, ioe);
throw (new Exception("Can not open History File"));
}
JobHistoryParser parser = new JobHistoryParser(in);
final EventReader realReader = new EventReader(in);
EventReader reader = Mockito.mock(EventReader.class);
if (numMaps == numSuccessfulMaps) {
reader = realReader;
} else {
final AtomicInteger numFinishedEvents = new AtomicInteger(0); // Hack!
Mockito.when(reader.getNextEvent())
.thenAnswer(
new Answer<HistoryEvent>() {
public HistoryEvent answer(InvocationOnMock invocation) throws IOException {
HistoryEvent event = realReader.getNextEvent();
if (event instanceof TaskFinishedEvent) {
numFinishedEvents.incrementAndGet();
}
if (numFinishedEvents.get() <= numSuccessfulMaps) {
return event;
} else {
throw new IOException("test");
}
}
});
}
jobInfo = parser.parse(reader);
numFinishedMaps = computeFinishedMaps(jobInfo, numMaps, numSuccessfulMaps);
if (numFinishedMaps != numMaps) {
Exception parseException = parser.getParseException();
Assert.assertNotNull("Didn't get expected parse exception", parseException);
}
}
Assert.assertEquals(
"Incorrect username ", System.getProperty("user.name"), jobInfo.getUsername());
Assert.assertEquals("Incorrect jobName ", "test", jobInfo.getJobname());
Assert.assertEquals("Incorrect queuename ", "default", jobInfo.getJobQueueName());
Assert.assertEquals("incorrect conf path", "test", jobInfo.getJobConfPath());
Assert.assertEquals("incorrect finishedMap ", numSuccessfulMaps, numFinishedMaps);
Assert.assertEquals("incorrect finishedReduces ", numReduces, jobInfo.getFinishedReduces());
Assert.assertEquals("incorrect uberized ", job.isUber(), jobInfo.getUberized());
Map<TaskID, TaskInfo> allTasks = jobInfo.getAllTasks();
int totalTasks = allTasks.size();
Assert.assertEquals("total number of tasks is incorrect ", (numMaps + numReduces), totalTasks);
// Verify aminfo
Assert.assertEquals(1, jobInfo.getAMInfos().size());
Assert.assertEquals(MRApp.NM_HOST, jobInfo.getAMInfos().get(0).getNodeManagerHost());
AMInfo amInfo = jobInfo.getAMInfos().get(0);
Assert.assertEquals(MRApp.NM_PORT, amInfo.getNodeManagerPort());
Assert.assertEquals(MRApp.NM_HTTP_PORT, amInfo.getNodeManagerHttpPort());
Assert.assertEquals(1, amInfo.getAppAttemptId().getAttemptId());
Assert.assertEquals(
amInfo.getAppAttemptId(), amInfo.getContainerId().getApplicationAttemptId());
Assert.assertTrue(
amInfo.getStartTime() <= System.currentTimeMillis()
&& amInfo.getStartTime() >= amStartTimeEst);
ContainerId fakeCid = BuilderUtils.newContainerId(-1, -1, -1, -1);
// Assert at taskAttempt level
for (TaskInfo taskInfo : allTasks.values()) {
int taskAttemptCount = taskInfo.getAllTaskAttempts().size();
Assert.assertEquals("total number of task attempts ", 1, taskAttemptCount);
TaskAttemptInfo taInfo = taskInfo.getAllTaskAttempts().values().iterator().next();
Assert.assertNotNull(taInfo.getContainerId());
// Verify the wrong ctor is not being used. Remove after mrv1 is removed.
Assert.assertFalse(taInfo.getContainerId().equals(fakeCid));
}
// Deep compare Job and JobInfo
for (Task task : job.getTasks().values()) {
TaskInfo taskInfo = allTasks.get(TypeConverter.fromYarn(task.getID()));
Assert.assertNotNull("TaskInfo not found", taskInfo);
for (TaskAttempt taskAttempt : task.getAttempts().values()) {
TaskAttemptInfo taskAttemptInfo =
taskInfo.getAllTaskAttempts().get(TypeConverter.fromYarn((taskAttempt.getID())));
Assert.assertNotNull("TaskAttemptInfo not found", taskAttemptInfo);
Assert.assertEquals(
"Incorrect shuffle port for task attempt",
taskAttempt.getShufflePort(),
taskAttemptInfo.getShufflePort());
if (numMaps == numSuccessfulMaps) {
Assert.assertEquals(MRApp.NM_HOST, taskAttemptInfo.getHostname());
Assert.assertEquals(MRApp.NM_PORT, taskAttemptInfo.getPort());
// Verify rack-name
Assert.assertEquals("rack-name is incorrect", taskAttemptInfo.getRackname(), RACK_NAME);
}
}
}
}
/**
* Go through a job and update the member variables with counts for information to output in the
* page.
*
* @param job the job to get counts for.
*/
private void countTasksAndAttempts(Job job) {
final Map<TaskId, Task> tasks = job.getTasks();
if (tasks == null) {
return;
}
for (Task task : tasks.values()) {
switch (task.getType()) {
case MAP:
// Task counts
switch (task.getState()) {
case RUNNING:
++this.mapsRunning;
break;
case SCHEDULED:
++this.mapsPending;
break;
default:
break;
}
break;
case REDUCE:
// Task counts
switch (task.getState()) {
case RUNNING:
++this.reducesRunning;
break;
case SCHEDULED:
++this.reducesPending;
break;
default:
break;
}
break;
default:
throw new IllegalStateException("Task type is neither map nor reduce: " + task.getType());
}
// Attempts counts
Map<TaskAttemptId, TaskAttempt> attempts = task.getAttempts();
int newAttempts, running, successful, failed, killed;
for (TaskAttempt attempt : attempts.values()) {
newAttempts = 0;
running = 0;
successful = 0;
failed = 0;
killed = 0;
if (TaskAttemptStateUI.NEW.correspondsTo(attempt.getState())) {
++newAttempts;
} else if (TaskAttemptStateUI.RUNNING.correspondsTo(attempt.getState())) {
++running;
} else if (TaskAttemptStateUI.SUCCESSFUL.correspondsTo(attempt.getState())) {
++successful;
} else if (TaskAttemptStateUI.FAILED.correspondsTo(attempt.getState())) {
++failed;
} else if (TaskAttemptStateUI.KILLED.correspondsTo(attempt.getState())) {
++killed;
}
switch (task.getType()) {
case MAP:
this.newMapAttempts += newAttempts;
this.runningMapAttempts += running;
this.successfulMapAttempts += successful;
this.failedMapAttempts += failed;
this.killedMapAttempts += killed;
break;
case REDUCE:
this.newReduceAttempts += newAttempts;
this.runningReduceAttempts += running;
this.successfulReduceAttempts += successful;
this.failedReduceAttempts += failed;
this.killedReduceAttempts += killed;
break;
default:
throw new IllegalStateException("Task type neither map nor reduce: " + task.getType());
}
}
}
}
@Override
public void updateAttempt(TaskAttemptStatus status, long timestamp) {
super.updateAttempt(status, timestamp);
TaskAttemptId attemptID = status.id;
TaskId taskID = attemptID.getTaskId();
JobId jobID = taskID.getJobId();
Job job = context.getJob(jobID);
if (job == null) {
return;
}
Task task = job.getTask(taskID);
if (task == null) {
return;
}
TaskAttempt taskAttempt = task.getAttempt(attemptID);
if (taskAttempt == null) {
return;
}
Long boxedStart = startTimes.get(attemptID);
long start = boxedStart == null ? Long.MIN_VALUE : boxedStart;
// We need to do two things.
// 1: If this is a completion, we accumulate statistics in the superclass
// 2: If this is not a completion, we learn more about it.
// This is not a completion, but we're cooking.
//
if (taskAttempt.getState() == TaskAttemptState.RUNNING) {
// See if this task is already in the registry
AtomicLong estimateContainer = attemptRuntimeEstimates.get(taskAttempt);
AtomicLong estimateVarianceContainer = attemptRuntimeEstimateVariances.get(taskAttempt);
if (estimateContainer == null) {
if (attemptRuntimeEstimates.get(taskAttempt) == null) {
attemptRuntimeEstimates.put(taskAttempt, new AtomicLong());
estimateContainer = attemptRuntimeEstimates.get(taskAttempt);
}
}
if (estimateVarianceContainer == null) {
attemptRuntimeEstimateVariances.putIfAbsent(taskAttempt, new AtomicLong());
estimateVarianceContainer = attemptRuntimeEstimateVariances.get(taskAttempt);
}
long estimate = -1;
long varianceEstimate = -1;
// This code assumes that we'll never consider starting a third
// speculative task attempt if two are already running for this task
if (start > 0 && timestamp > start) {
estimate = (long) ((timestamp - start) / Math.max(0.0001, status.progress));
varianceEstimate = (long) (estimate * status.progress / 10);
}
if (estimateContainer != null) {
estimateContainer.set(estimate);
}
if (estimateVarianceContainer != null) {
estimateVarianceContainer.set(varianceEstimate);
}
}
}
